How Memory Works
Or, a hypersimplistic intro to neural networks
👋 Hey there! Welcome to a new edition of The Sunday Wisdom! My name is Abhishek. I read a lot of books, think a lot of things, and this is where I dump my notes and (so called) learnings.
I mostly write to educate myself; this is kind of my Feynman Technique in action. But if you like my writing, I would say this little hobby of mine just became a bit more purposeful. The latest two editions are always free. The rest are available to paid subscribers. Now… time for the mandatory plug!
But seriously, if you are facing trouble completing the subscription above, you can alternatively make a pledge on Patreon (if you are sooo keen). Pledge whatever amount you want and I’ll unlock paid posts for you. Deal?
Alright! On to this week’s essay. It’s about 3,050 words.
Before I begin, I must say that you can expect to look forward to a couple of neuroscience-y essays in the next few weeks, because that’s mostly what I’ve been reading (and have been really interested in) lately.
I’ve been thinking for a while about going deep, really deep, into a particular subject, and “the brain” kept coming up in my mind (which is actually controlled by my brain) over and over again.
So yes, expect a couple of brain-y essays, with occasional bouts of other subjects such as psychology, philosophy, business, etc. Alright! Let’s go!
Q: What is memory really, and how does it even work?
First, let me tell you the story of H.M. who was one of the most famous patients in medical history.
When H.M. (his real name was Henry Molaison, but scientists shrouded his identity until his death in 2008) was seven years old, he was hit by a bicycle and landed hard on his head. Soon afterward, he developed seizures and started blacking out. At sixteen, H.M. had his first grand mal seizure, the kind that affects the entire brain. Soon, he was losing consciousness up to ten times a day.
By the time he turned twenty-seven, H.M. was desperate. Anticonvulsive drugs (the ones used to treat epileptic seizures) hadn’t helped. He was smart, but couldn’t hold a job. He still lived with his parents. But H.M. wanted a normal existence, so he sought help from a physician whose tolerance for experimentation outweighed his fear of malpractice.
The doctor concluded that an area of the brain called the hippocampus was the main culprit behind his seizures. When he proposed cutting into H.M.’s head, lifting up the front portion of his brain, and, with a small straw, sucking out the hippocampus and some surrounding tissue from the interior of his skull, H.M. agreed. Like I said, he was desperate.
The surgery occurred in 1953, and as H.M. healed, his seizures slowed. Almost immediately, however, it became clear that his brain had been radically altered.
H.M. knew his name and that his mother was from Ireland. He could remember the 1929 stock market crash and news reports about the invasion of Normandy. But almost everything that came after his surgery — all the memories and experiences — had been erased.
When a doctor began testing H.M.’s memory by showing him playing cards and lists of numbers, he discovered that H.M. couldn’t retain any new information for more than twenty seconds or so.
From the day of his surgery until his death in 2008, every person H.M. met, every song he heard, every room he entered, was a completely fresh experience. His brain was frozen in time. Kind of similar to what happened to the character of Guy Pearce in Christopher Nolan’s Memento.
He introduced himself to his doctors and nurses over and over, dozens of times each day. And each day, he was befuddled by the fact that someone could change the television channel by pointing a black rectangle of plastic at the screen.
Even though we know a good deal about how memories are assembled and how and where they are stored, we don’t exactly know why we keep some and not others.
It clearly has little to do with actual value or utility. For example, I can remember the entire lineup of the 2003 Indian Cricket World Cup team — something that has been of no importance to me since 2003 and wasn’t actually very useful even then — and yet I cannot recollect the number of my mother’s cell phone (whom I call everyday), or where I left my phone just five minutes ago, or what was the third of three things I was supposed to get at the shop, or any of a great many other things that are unquestionably more urgent and important than remembering the players of a sports team.
One particularly important distinction in storage is between short-term memory and long-term memory. With the former, you look up a phone number, sprint across the room convinced you’re about to forget it, punch in the number, and then it’s gone forever. Short-term memory is your brain’s equivalent of juggling some balls in the air for 30 seconds.
In contrast, long-term memory refers to remembering what you had for dinner last night, the name of the leader of your nation, how many grandchildren your grandfather has, or where you went to college. But that’s not all, the interesting part is that there’s a specialised subset of long-term memory called remote memory.
Remote memories are ones stretching back to your childhood — the name of your village, your native language, the smell of your grandmother’s cooking. They appear to be stored in some sort of archival way in your brain, separate from more recent long-term memories. Often, in patients with dementia that devastates most long-term memory, the more remote facets can remain intact.
Another important distinction in memory is that between explicit memory (also known as declarative memory) and implicit memory, which includes an important subtype called procedural memory.
Explicit memory concerns facts and events, along with your conscious awareness of knowing them: I am a mammal, today is Sunday, my father sported a moustache. Things like that. In contrast, implicit procedural memories are about skills and habits, about knowing how to do things, even without having to think consciously about them: shifting the gears on a car, riding a bicycle, playing the guitar, or typing at a keyboard.
Memories can be transferred between explicit and implicit forms of storage. For example, when you are learning a new language, each time you form a sentence, you must consciously remember what to do — think of the correct word, think of its tense, consider how the words would be arranged in the sentence, see if there’s some kind of liaison (such as in French), etc.
And then, with enough practice, one day you suddenly realise that you just formed a sentence smoothly without even thinking about it. In technical terms, what just happened is you did it with implicit, rather than explicit, memory. If you are learning to play the guitar, this happens when you realise for the first time as if your fingers have a mind of their own.
The funny thing is that memory can be dramatically disrupted if you force something that’s implicit into explicit channels. Here’s an example that will finally make reading this essay worth your while — how to make neurobiology work to your competitive advantage in sports.
Suppose you’re playing cricket against a team and one of their batsmen is hitting your balls out of the park with ease. Wait until they hit a particularly impressive shot, and then offer a genuine compliment like, “Wow, that was an amazing shot! I couldn’t have hit it like that in a million years. How did you manage to time it so perfectly? Did you pick up any cues from my bowling action or the field setting?”
Then, follow up with specific questions about their technique, such as “Do you keep your eyes on the ball till the last moment? How do you position your feet before the ball is bowled?” Basically, not only give a genuine compliment, but also give it in excruciating details.
Do it right, and the next time that shot is called for, your opponent will make the mistake of thinking about it explicitly, and the hit won’t be anywhere near as effective.
You can’t think and hit (or rather, do anything) at the same time. As Andre Agassi used to say to himself on the court, “Don’t think, Andre. Turn off your mind.” As Tom Cruise says in Top Gun: Maverick, “You think up there, you’re dead.”
Just imagine descending a flight of stairs in an explicit manner, something you haven’t done since you were two years old — okay, bend my left knee and roll the weight of my toes forward while shifting my right hip up slightly — and down you go. You would topple.
Now, just as there are different types of memories, there are different areas of the brain involved in memory storage and retrieval. One critical site is the cortex, the vast and convoluted surface of the brain. Another is a region tucked just underneath part of the cortex, called the hippocampus; the part H.M. was missing.
Both of these are regions vital to memory. For example, it is the hippocampus and cortex that are gradually damaged in Alzheimer’s disease. You can think of the cortex as your hard drive, where memories are stored, and your hippocampus as the keyboard, the means by which you access memories in the cortex. With his hippocampus removed, H.M. didn’t have this keyboard to retrieve memories from his hard drive.
But that’s not all. There are additional brain regions relevant to a different kind of memory. These are structures that regulate body movements, such as the cerebellum. But what does it have to do with memory, you ask? Well, they appear to be relevant to implicit procedural memory, the type you need to perform reflexive, motor actions without even consciously thinking about them, where, so to speak, your body remembers how to do something before you do (often referred to as habit or muscle memory).
This is why, despite this profound amnesia, H.M. could still do basic stuff, such as brushing his teeth, or going to the bathroom. Not only that. Give him some mechanical puzzle to master day after day, and he could learn to put it together at the same speed as anyone else, while steadfastly denying each time that he had ever seen it before. Although his hippocampus and explicit memory was shot, the rest of his brain was intact, and so was his ability to acquire procedural memories.
Alright! With the basic stuff taken care of, let’s shift gears and go a little bit deeper, literally. Let’s try to understand what’s going on at the level of clusters of neurones within the cortex and hippocampus.
A long-standing belief among many who studied the cortex was that each individual cortical neurone would, in effect, turn out to have only a single task, a single fact that it knew. This was prompted by some staggeringly important work done in the 1960s by David Hubel and Torstein Wiesel of Harvard on the visual cortex.
They used microelectrodes to record the electrical activity of individual neurones in the visual cortex of anesthetised cats, while presenting visual stimuli such as light, lines, and shapes to the animals.
Hubel and Wiesel’s research demonstrated that the processing of visual information in the brain is hierarchical, with simpler features such as lines and edges processed in lower levels of the visual cortex, and more complex features such as objects and faces processed in higher levels.
While one level responded to a single dot of light, another level would respond to a sequence of adjacent dots, i.e., a straight line of light, and this led people to believe that there would be another level, the third level, where each neurone responded to a particular collection of lines, and a fourth and fifth level, all the way up until, at the umpteenth level, there would be a neurone that responded to one thing and one thing only, namely your grandmother’s face at a particular angle (and next to it would be a neurone that recognised her face at a slightly different angle, and then the next one… and so on).
People went out looking for what were actually called grandmother neurones — neurones way up in the levels of the cortex that “knew” one thing and one thing only, namely a complexly integrated bit of sensory stimulation. With time, it became apparent that there could be very few such neurones in the cortex, because you simply don’t have enough neurones to allow each one to be so narrow-minded and overspecialised.
Rather than memory and information being stored in single neurones, they are stored in the patterns of excitation of vast arrays of neurones — in trendy jargon, memories are stored in neural networks.
Let me illustrate how they work with a wildly simplified example: Suppose there are three neurones. A neurone called “Gauguin” knows how to recognise Gauguin paintings; while a neurone called “Van Gogh” knows about Van Gogh paintings; and a third neurone called “Monet” can recognise… you guessed it, Monet paintings.
Now, the grandmother neurones (neurones that know one thing and one thing only) send information to a second level in this network, comprising of neurones A, B, C, D, E (sorry I couldn’t find fancier names). For example, Gauguin talks to A, B, and C; Van Gogh talks to B, C, and D; Monet talks to C, D, and E.
Now, humour me for a second. What “knowledge” does neurone A have? It gets information only about Gauguin paintings, which is a grandmotherly neurone. Similarly, E gets information only about Monet paintings.
But C is interesting. Neurone C has access to all three grandmother neurones. So, what does C know about? C knows about “Impressionism”, the features that these three painters have in common. Basically, C is the neurone that says, “I can’t tell you the painter, certainly not the painting, but I can tell you this much: it’s one of those Impressionists.”
C has knowledge that does not come from any single informational input, but emerges from the convergence of information feeding into it. Neurones B and D are also Impressionism neurones but, because they have fewer examples to work with, they’re just not as good as neurone C. Most neurones in your cortex process memory like neurones B through D, not like A or E.
We take advantage of such convergent networks all the time. The best example is whenever you are trying to think of something that is almost… almost there.
Suppose you’re trying to remember the name of a painter: That guy, what’s his name… he was that Spanish guy with a killer moustache (activating your “Spanish guy” and “killer moustache” network). He painted those complex and enigmatic compositions. It isn’t Goya (two more networks are pulled in). I watched all those elaborate analyses of his paintings on YouTube… oh, I even saw his most famous painting when I visited that museum in Madrid. If only I could remember the name of that museum, or maybe the name of that painting, I bet I could remember his name.
With enough of those nets being activated, you finally stumble into the one fact that is at the intersection of all of them: Diego Velázquez, the equivalent of a neurone C.
That’s a rough approximation of how a neural network operates, and neuroscientists have come to think of both learning and storing of memories as involving the “strengthening” of some branches rather than others of a network.
How does such strengthening occur? For that, we have to go one more level deeper in order to consider the tiny gaps between the thready branches of two neurones, called synapses.
You see, when a neurone has heard some fabulous gossip and wants to pass it on, a wave of electrical excitation sweeps over it, which triggers the release of chemical messengers — neurotransmitters — that float across the synapse and excite the next neurone.
There are dozens, probably hundreds, of different kinds of neurotransmitters, and synapses in the hippocampus and cortex disproportionately make use of what is probably the most excitatory neurotransmitter there is, something called glutamate.
Besides being superexcitatory, “glutamatergic” synapses have two properties that are critical to memory. The first is that these synapses are nonlinear in their function. What does this mean? In a run-of-the-mill synapse, when a little bit of neurotransmitter comes out of the first neurone, it causes the second neurone to get a little excited. If a smidgen more neurotransmitter is released, there is a smidgen more excitation, and so on.
In glutamatergic synapses however, when some glutamate is released, nothing happens. When a larger amount is released, nothing happens. It isn’t until a certain threshold of glutamate concentration is passed that, suddenly, all hell breaks loose in the second neurone, and there is a massive wave of excitation. This is what learning something new feels like.
You are reading this essay, trying to make sense of it, trying to form a mental model in your head, and nothing is happening. Even though you are reading these lines, they are simply going in and out of your brain without being processed. Finally, when you read it for the hundredth time, a lightbulb goes on, “Aha!” and you get it. On a simplistic level, when you finally get it, that nonlinear threshold of glutamate excitation has just been reached.
The second feature is even more important. Under the right conditions, when a synapse has just had a sufficient number of superexcitatory glutamate-driven ahas, something interesting happens.
The synapse becomes persistently more excitable, so that next time it takes less of an excitatory signal to get that aha. That’s why, once you learn the fundamentals of a subject, it becomes relatively easier to make sense of the complex topics in that subject. Technically, the synapses get “potentiated” or strengthened.
The most amazing thing is that this strengthening of the synapse can persist for a long time. In other words, once you really learn something, after hitting that aha moment, the knowledge stays on in your memory for a long time.
So… that’s it! This is broadly all you need to know about how your brain remembers anniversaries, and sports statistics, and the colour of someone’s eyes, and how to dance.
But… before I end, back to H.M. one more time.
While, H.M couldn’t identify his psychiatrist who he had been seeing for years, he could always recognise himself in the mirror, but was often astounded at how old he had become. Occasionally, and mysteriously, he was able to lay down just a few memories. He could recall that John Glenn was an astronaut and Lee Harvey Oswald an assassin (though he couldn’t recall whom Oswald had assassinated) and learned the address and layout of his new house when he moved. But beyond that he was locked in an eternal present that he could never understand.
Poor Henry Molaison’s plight was the first scientific intimation that the hippocampus has a central role in laying down memories. But what scientists learned the most from Molaison was not so much about how memory works, but how difficult it is to really understand how memory works.
Today I Learned
In the 1600s, Peruvian fishermen noticed something odd. During certain periods, the sea water turned warmer, and their catch of fish dwindled. This affected their livelihood. But they didn’t really know what was happening or why.
Over time this phenomenon started came to be known kas El Niño, which literally means ‘The Boy’ in Spanish — no idea why. (Yes, there’s something call La Niña, ‘The Girl’, which is an oceanic and atmospheric phenomenon that is the colder counterpart of El Niño.)
But it took another 300 years before scientists explained to us what was really going on. It was found to be linked to something called the Southern Oscillation over the tropical Pacific Ocean.
So this is what happens normally. The trade winds blowing from east to west push the warm surface waters of the western Pacific toward Indonesia and Australia, while cooler waters from the depths of the ocean rise up to replace them. This process is known as upwelling, and it brings nutrient-rich cold water to the surface, which supports the growth of plankton and fish.
But every once in a while, the cycle changes. The trade winds weaken or even reverse direction, allowing the warm surface waters to flow back toward the east, towards the coast of South America. This causes the water temperatures in the eastern and central Pacific to rise. The temperature rises anywhere from 1°C to 3°C compared to normal. And the fish disappear. Scientists called this the El Niño Southern Oscillation (ENSO).
Now, even though it appears that it only affects Peru, no, El Niño affects everyone.
There are ripple effect across the globe. Or in climate science terms, it has a teleconnection — geographically distinct areas are affected by a singular weather pattern. This includes India.
See, when El Niño arrives, the complex dance involving atmospheric pressure, winds, and sea temperatures affects the monsoons. This is why India has been hotter than normal for the last couple of weeks, with no sign of rain.
The monsoon winds, which normally blow from the southwest, can weaken or even reverse direction when El Niño rears its head, causing a decrease in rainfall over parts of India, particularly in the northwest region. Depending on its severity, this can lead to droughts, crop failures, and other problems. In fact, 50% of droughts in India over the past 130 years have been linked to an El Niño.
Timeless Insight
We can’t derive Laws of History that predict the future the way we can with, say, a law of physics that carries predictive capability under stated conditions. (i.e., if I drop a ceramic coffee cup from my table, it will shatter.)
We can only merely deduce some tendencies of human nature and generate some reasonable expectation that if X happens, Y is somewhat likely to follow. But viewing the process of human history as possessing the regularity of a solar system is folly.
History is a unique process — it only gets run once.
What I’m Reading
The best moments in our lives, are not the passive, receptive, relaxing times — although such experiences can also be enjoyable, if we have worked hard to attain them. The best moments usually occur when a person’s body or mind is stretched to its limits in a voluntary effort to accomplish something difficult and worthwhile.
— Mihaly Csikszentmihalyi, Flow: The Psychology of Optimal Experience
Tiny Thought
Human beings have an infinite capacity for taking things for granted.
Before You Go…
Thanks so much for reading! Send me ideas, questions, reading recs. You can write to abhishek@coffeeandjunk.com, reply to this email, or use the comments. And… if you feel like I’ve done a great job writing this piece, be generous and buy me a few cups. ☕️
Until next Sunday,
Abhishek 👋
PS: All typos are intentional and I take no responsibility whatsoever!